Event
DEFLECT: Improving Latency Robustness of Asynchronous VLA with Temporal Counterfactual Preference Learning
1 reports · 1 sources
Yixiang Zhu et al. propose DEFLECT, an offline post-training framework that converts the temporal discrepancy between prediction and execution into preference supervision, enabling vision-language-action (VLA) policies to produce actions aligned with the state at execution time even when inputs are already stale. In evaluations on Kinetix, LIBERO, and three real-robot tasks, DEFLECT improved latency robustness across the tested VLA policies and settings: at the longest supported latency, it achieved a 4.2 percentage point higher success rate than VLASH on Kinetix, and on LIBERO with GR00T N1.7 it achieved a 22.7 percentage point higher success rate than naive asynchronous execution.
Report timeline
- DEFLECT: Temporal Counterfactual Preference Learning for Delay-Robust Asynchronous VLAs
DEFLECT is an offline post-training framework that turns the temporal gap between prediction and execution into preference supervision, enabling policies to produce actions aligned with the execution-time state even when their inputs are stale.
arXiv cs.RO ·